About two hundred thermal indicators exist and yield divergent assessments of heat stress impacts and mitigation. Thus, examining how these indicators respond to various meteorological variables and exploring the implications for their practical use is imperative. Using a correlation analysis, we cluster common indicators into three types: 1) human energy budget models, 2) integrated weather indices, and 3) thermal perception indicators. Distinct extreme hot conditions are identified differently by the various clusters of indicators: human energy budget models are more responsive to micro-scale variation in wind and radiation; while integrated weather indices mainly capture synoptic moist heat extremes. These biophysical indicators also do not concur with a metamodel of thermal perception, developed here using a meta-analysis of coefficients in existing thermal sensation vote equations. The developed thermal perception metamodel is more sensitive to radiation fluxes than other thermal stress indicators. It implies that humans’ thermal sensation may underestimate humid heat stress at nighttime, which can pose a significant risk to human health in hot, humid cities such as Chennai (India) and Dakar (Senegal) and across the Global South. These findings deepen our understanding of heat stress variability on humans and provide a framework for selecting suitable indicators in future applications.

In June 2023 Swiss people voted a new climate law that set a net-zero emission goal to be reached by 2050 via a full energetic transition from fossil fuels to renewables. The country’s Energy Strategy estimates that 7% (4.3 TWh) of future total renewable energy will be supplied by wind turbines, which requires an increase in the number of installed devices from the 37 currently operating to 760. Such an objective presents numerous challenges as available space is limited by technical restrictions, the country’s complex terrain, and competition with other types of land use.Thanks to qualities like small size and weight, low noise emission levels, and the ability to operate with winds blowing from any direction at relatively low speed (> 2 m/s), vertical axis wind turbines (VAWTs) installed in urban areas are an attractive alternative to overcome the issues associated with large wind farms. Despite this, the potential for wind energy micro-generation in complex urban settings remains largely unexplored.Private households use one third of all energy consumed in Switzerland, and residential renewable energy generation currently consists almost exclusively of photovoltaic (PV) panels which, in 2021, represented 78% of all solar systems operating in the country. No similar statistics are available for residential wind energy generation. Even in the scientific literature, current understanding of the interaction between wind and urban areas is limited, and the knowledge about urban wind resources is markedly inadequate to address the challenges posed by climate change to both local and global energy sectors.Here we use use the Weather Research and Forecast (WRF) model to simulate mean near-surface wind speed over the cities of Lausanne and Geneva to assess the potential for wind energy generation. We perform simulations at 300 m grid spacing and across 85 vertical model levels, with hourly output interval throughout one entire year to identify diurnal and seasonal wind speed trends. We then use power curves of select VAWTs to translate mean wind speed data into potential electrical output maps and time series, over all model cells classified as urban.  Our results show that mean wind speed is generally higher in Lausanne than in Geneva, especially at nighttime. Diurnal cycles evolve markedly differently between the two cities, although differences are at times minimized due to seasonal changes. The average potential for wind energy harvesting using VAWTs in urban environments varies with turbine size and geographical area. The average daily total energy generation potential is one order of magnitude greater in Lausanne compared to Geneva. In Lausanne, top generation is expected during the nighttime across most months, allowing for a good integration of photovoltaic generation. The opposite happens in Geneva where already lower peak wind speed, and associated energy generation, always culminate during the afternoon.This research highlights the potential for urban wind energy micro-generation, drawing attention to the role of regional differences and the need and the importance of numerical simulations for quantitative assessments at the city and regional scales.

Urban-induced changes in local microclimate, such as the urban heat island effect and air pollution, are known to vary with city size,  leading to power law or logarithmic relations between average climate variables and city-scale quantities (e.g., total population or area). However, these approaches suffer from biases related to the choice of city boundaries and they neglect intra-urban variations of city properties. In this study we use high-resolution data of urban temperatures and annual concentrations of particulate matter together with population density and street network properties and show that their marginal and joint probability distributions follow universal finite-size scaling functions. These results extend previous findings on city-scale relations, offering a novel description of intra-urban fluctuations of climate characteristics.

Cities are increasingly adopting blue-green infrastructure (BGI) to address the dual challenges of extreme rainfall and rising temperatures driven by climate change. While the potential of BGI for urban stormwater management is well-studied, the cooling effect of stormwater-focused BGI remains underexplored. This study investigates the heat mitigation potential of three stormwater BGI elements, bioretention cells, porous pavements, and detention ponds, within three urban street canyons in a Swiss town near Zurich. The Urban Tethys-Chloris (UT&C) microclimate model was modified to explicitly represent stormwater BGI and assess their influence on the Universal Thermal Climate Index (UTCI) at 2 meters above the ground. Simulations were conducted under both historical climate and a future climate projection, including a sensitivity analysis of soil types. Results show that stormwater BGI reduce median UTCI by 0.2 to 0.5 °C, with peak reductions reaching up to 2.7 °C. However, their effectiveness depends on the type of BGI, the surface it replaces, and the availability of water. Soil properties were found to significantly influence the cooling effect of bioretention cells, with finer-textured soils achieving higher soil moisture levels and greater reductions in UTCI. A trade-off between cooling benefits and stormwater management also emerged: sandy soils favor infiltration but dry quickly, limiting cooling, while clay-rich soils limit infiltration but retain moisture and sustain evaporative cooling, even under future climate conditions with longer dry spells. These findings highlight the importance of integrating hydrological and thermal considerations into BGI design and suggest an integrated design that balance both objectives.

Urban heat stress is a growing public health concern, particularly for soft mobility users such as cyclists and pedestrians. Understanding how urban and green infrastructure design, together with local micrometeorology, influences cycling behavior, is essential for developing effective heat mitigation strategies and promote sustainable urban mobility.Using the Brussels Capital Region (BCR) as a case study, this research utilizes data collected by Brussels Mobility, which deployed 18 monitoring stations to continuously record bicycle count and speed at 15-minute intervals. For the year 2023, the data was aggregated into hourly intervals, and Mean Radiant Temperature (Tmrt), Universal Thermal Climate Index (UTCI), and Physiologically Equivalent Temperature (PET) were computed for a 25 m × 25 m area surrounding each monitoring station. Additionally, the stations were classified as either green or non-green (26°C UTCI; >23°C PET). In warmer months, greenness also influenced cycling behavior, with a higher count ratio, especially on weekends and in the mornings. However, this effect was not observed when PET exceeded 35°C.Overall, thermal comfort, greenness and time of day have shown to affect cycling behavior. These findings highlight the need for further research to examine the role of greenery in cycling behavior, particularly given the limited route options available to cyclists.

Accurate quantum mechanical (QM) simulations are central to understanding the electronic structure and properties of molecules and materials. Electronic structure methods solve the Schrödinger equation for the electronic Hamiltonian, from which all ground and excited state properties can be derived. However, their steep computational cost limits applications to small systems or short timescales. Machine learning (ML) offers a way forward by creating surrogate models that map structure to properties at much lower cost. Early ML approaches focus on specific observables, energies, forces, charges, dipole moments, polarizabilities, rather than the underlying electronic structure. While effective within their training domain, such models lack transferability and cannot predict properties beyond those in the training data. A more general solution is to learn fundamental electronic quantities, such as electron densities, density matrices, wave functions, or effective single-particle Hamiltonians, which provide access to many properties through inexpensive postprocessing. This thesis focuses on learning an effective single-particle Hamiltonian.

Accurate Hamiltonians require large basis sets, producing high-dimensional matrices that make direct learning difficult. To balance accuracy and efficiency, we introduce an indirect learning framework. Instead of using matrix elements as final targets, the Hamiltonian is treated as an intermediate representation, while learning targets are derived properties such as orbital energies, charges, or observables computed in either the model basis or a larger reference basis. The model remains parametrized in a compact minimal basis, reducing complexity while still guided by information from more complete calculations. This hybrid design improves efficiency without sacrificing accuracy and preserves access to a wide range of properties through postprocessing. Using automatic differentiation, we optimize the effective Hamiltonian to reproduce observables from either the same or larger basis. Coupled with the Tammâ Dancoff approximation, ML-predicted Hamiltonians can predict singlet excited states across molecules. The models generalize well to unseen, larger systems while being orders of magnitude faster than reference methods, enabling applications such as computing spectral densities from molecular dynamics.

We extend this framework by interfacing ML Hamiltonians with PySCFAD, an auto-differentiable electronic structure code supporting density matrix construction and linear response calculations. This greatly expands the scope of indirect models, providing access to many observables without reimplementing differentiable routines. We analyze how design choices, such as adding physical constraints or basis set parametrization, affect accuracy and transferability. Well-regularized models extrapolate reliably to larger molecules, and for properties like dipole moments and polarizabilities, Hamiltonian-based models outperform property-specific ones. We also extend the framework to periodic systems for predicting band energies.

This thesis shows that ML Hamiltonians offer a powerful and generalizable bridge between QM and ML. By targeting an operator central to electronic structure rather than isolated properties, these models deliver efficient surrogates capable of predicting diverse observables with high accuracy. This work points toward hybrid MLâ QM approaches that unify accuracy and efficiency.

In recent years, machine learning-based generative models have emerged as promising tools for de novo molecular design, offering a data-driven means of proposing novel small molecules with tailored properties. These models are increasingly applied in drug discovery, with generated molecules experimentally validated and currently in clinical trials. However, two persistent challenges limit translational practicality: The ability to efficiently tailor the generation towards (computationally predicted) property-optimal molecules (sample efficiency), and the requirement that generated molecules are synthetically accessible using plausible, known chemistry. Focusing on language-based molecular generative models, this thesis makes contributions toward both challenges.

Part I starts by investigating design features that can enhance the sample efficiency of reinforcement learning-based molecular optimization, and the imposed trade-offs. Using these insights, a molecular generative framework is proposed that can perform optimization under highly constrained computational budgets. The model is applied across multi-parameter optimization tasks spanning drug discovery and adjacent fields such as functional materials design. Part II builds on this framework and addresses increasingly granular definitions of synthesizability by coupling machine learning-based retrosynthesis models that output predicted synthesis routes given input molecules. By comparing the synthesizability of generated molecules as guided by heuristic scores and retrosynthesis models, our analysis highlights out-of-distribution limitations of the former and demonstrates a practical advantage for considering explicit synthesis routes. The framework is subsequently extended to enable steerable and granular synthesizability control. Accordingly, generated molecules have associated synthesis routes incorporating specific chemical reagents, specific reactions, and avoiding other reactions. The ability to control reaction constraints also shows the potential to unify generative design and ultra-large-scale (> billion scale) virtual screening. Specifically, the framework can generate property-optimal molecules with exact matches in so called, "make-on-demand" molecular libraries, i.e., directly purchasable from a vendor, thus performing retrieval. The molecular generative models developed in this thesis lower the barrier for experimental validation and make contributions toward enhanced translational practicality.

Time, though ubiquitous in our daily experience, is a physical quantity not very well understood. It lies at the crux of efforts to unify the physical paradigms, being regarded as a constant background in some paradigms such as Newtonian mechanics or quantum mechanics, while completely internal to the objects in other paradigms such as general relativity. In quantum mechanics, specifically, the inability to find an operator for time as for spatial positions indicates the impossibility to find an observable which represents the external time \textit{coordinate}. Nevertheless, an operator can often be found to represent the internal time \textit{duration} of some certain process. The Eisenbud-Wigner-Smith (EWS) time scale is a famous example of such a time duration operator, which describes the single-particle scattering or ionization time scale.

The experimental determination of the EWS time scale for atomic photoionization has long been established and is a well-developed technique. With the assistance of ultrashort laser pulses, the relative time delay between photoionization from different states is measured, giving results in the attosecond ($10^{-18},$s) range. This technique essentially measures the difference in time \textit{coordinate} between two photoionization events, while the absolute time coordinate of a single event cannot be determined due to the difficulty of defining the time-zero for the event.

In this thesis, a complementary approach is used to access the absolute and internal time \textit{duration} of photoelectron emission without explicit time resolution. Being better suited for photoemission from dispersive bands in solid-state materials, this method relies on the interference between multiple photoemission channels, which gives rise to spin polarization of the photoelectrons. The kinetic energy derivative of spin polarization contains information on the phase shift accumulated in the photoemission process, which can then be used to estimate the EWS photoemission time scale.

Building upon previous results estimated by the analytical model, which are $\tau_{EWS}\approx26,$as for Cu(111) and $\tau_{EWS}\geq120,$as for BSCCO, three materials have been selected to extend the parameter space in correlation strength and dimensionality, namely the quasi-2-dimensional transition metal dichalcogenides (TMDCs) 1T-TiSe$_2$ and 1T-TiTe$_2$, and the quasi 1-dimensional CuTe. A thorough study with the angle-resolved photoemission spectroscopy (ARPES) has been carried out to investigate their electronic structures and charge ordered phases. Their spin polarizations were then measured with spin- and angle-resolved photoemission spectroscopy (SARPES) to get an estimate of their EWS photoemission time scales. 1T-TiSe$_2$ and 1T-TiTe$_2$ show EWS time scales around 150,as, whereas in CuTe the photoemission takes more than 200,as. Analysis of results obtained thus far demonstrates a pattern in the dimensionality, as well as the symmetry of the system under investigation, and also sheds lights on the understanding of electronic correlation.

Given their widespread occurrence in cold-region environments, snowdrifts have been the focus of extensive research aimed at accurately predicting their formation and spatial extent. Snowdrifts impact both human and natural systems, ranging from imposing critical snow loads on infrastructure to modifying the surface energy balance of snow-covered sea ice. Despite numerous modeling efforts in snow transport research, a standardized, widely accepted, and openly accessible approach for simulating snowdrift around obstacles remains lacking. This thesis presents snowBedFoam, an Eulerian-Lagrangian model designed to simulate aeolian snow transport and predict snowdrift formation around structures. Implemented within the OpenFOAM framework, it supports a broad set of turbulence models and boundary conditions, ensuring flexibility across diverse applications. To assess and validate its performance, the model was applied to three case studies representative of snow-covered environments: an Antarctic research station, alpine solar panels, and icebergs grounded in landfast sea ice. In each case, simulations over short timescales showed strong agreement with field observations on snow distribution. Concurrently, key drivers influencing snowdrift formation were evaluated. In the Antarctic station case, the sensitivity of snowdrifts to structural design, wind speed, and snow characteristics was analyzed; in the alpine solar farms, the effects of panel arrangement and mutual shielding were investigated; and for icebergs, the influence of size and wind conditions on snowdrift formation was specifically examined. Together, these three cases offer unique yet complementary perspectives, supporting a multifaceted understanding of snow transport dynamics around obstacles. Across all scenarios, simulations highlight the critical role of turbulent gusts and wind direction variability in accurately reproducing snow distribution. Snowbed cohesion also emerged as a critical parameter, reinforcing the need for precise surface property characterization, while snowdrift morphology was strongly influenced by fine surface geometry, with minor edges and contours influencing local deposition. In parallel, group effects between obstacles -such as wind shielding by adjacent structures- significantly altered local snow accumulation, often amplifying deposition on individual units. These insights point to several key directions for future research: integrating dynamic surface evolution, improving the parameterization of snow and aerodynamic processes, and expanding short-term empirical datasets to enable rigorous model validation. Addressing these challenges is essential for advancing a next-generation snowdrift model that combines accuracy and adaptability across a wide range of cold-region applications. Overall, these findings highlight the model's ability to capture the complex interplay between airflow, surface geometry, and snowpack properties that shape snowdrift formation. They also demonstrate its potential to inform infrastructure design, optimize renewable energy layouts, and support research in extreme polar environments - making snowBedFoam a valuable asset for both scientific and engineering applications.

As digital data become central to domains, ranging from personalized medicine and financial services to scientific research and national security, preserving its confidentiality during computation has emerged as a foundational challenge. Recent breaches illustrate the severe consequences of compromised information: privacy violations, identity theft, and loss of public trust. Traditional encryption protects data at rest and in transit, but leaves a critical gap during computation, when sensitive inputs are typically exposed. Secure computation techniques, notably fully homomorphic encryption (FHE) and garbled circuits (GC), close this gap by enabling meaningful computation over encrypted or distributed data without revealing the underlying inputs.

Despite remarkable progress, these protocols still suffer from high computational and communication overheads, limiting their practical deployment. This thesis targets a specific --- but broadly applicable --- scenario: the secure evaluation of Boolean functions using FHE and GC. While these protocols are not inherently restricted to Boolean logic, focusing on Boolean circuits allows us to leverage decades of progress in logic synthesis, a mature subfield of electronic design automation. By casting secure computation as a domain-specific logic synthesis problem --- one governed by cryptographic cost models rather than silicon constraints in the conventional context --- we develop novel circuit optimization techniques that bridge classical logic design and modern cryptography.

We make contributions across five technical directions: (i) For leveled FHE schemes, we formalize the trade-off between multiplicative depth (MD) and multiplicative complexity (MC), and introduce synthesis frameworks that jointly reduce both, enabling tighter cryptographic parameters and faster homomorphic evaluation. (ii) For fast-bootstrapping FHE schemes, represented by the torus FHE (TFHE) scheme, under a fixed-plaintext-space function-evaluation strategy, we develop symmetry- and negacyclicity-aware mapping strategies, combined with a multi-value programmable bootstrapping (MV-PBS)â aware lookup-table (LUT) synthesizer, to minimize PBS count. (iii) For TFHE-based, multi-plaintext-space function evaluation, we propose the first synthesis framework that strategically transitions between binary and large plaintext spaces, using XORs for linear logic and concentrating non-linear operations into compact LUTs. (iv) For GC, we introduce the XORâ OneHotâ inverter graph (X1G), a new intermediate representation that improves ciphertext efficiency and unlocks dedicated optimization passes. (v) Finally, we introduce joint multiplicative complexity (JMC), a garbling-cost-aware cost model that accounts for shared ownership in GC-based secure multi-party computation, and propose the first ownership-aware optimizer that reduces jointly evaluated non-linear gates.

Collectively, these results demonstrate two broad lessons: (1) representation matters --- the choice of intermediate form directly impacts achievable savings and the scope of optimizations; and (2) cost models must match cryptographic reality --- accurate modeling of protocol bottlenecks is essential to obtain meaningful efficiency gains. Beyond these technical findings, the thesis argues for stronger standardization and cross-layer interfaces between cryptographic libraries, compilers, and hardware, to ensure that optimizations at one layer remain aligned with advances at